International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
NMF-SVM Based CAD Tool for the Diagnosis
of Alzheimer’s Disease
Ms. Tejal A. Fuse1, Mr. Nikita D. Jayasignpure2 and Prof. Pragati D. Pawar3
1
Ms. Tejal A. Fuse Electronics & Telecomm. Department
J.D.I.E.T. Yavataml
2
Ms Nikita D. Jayasignpure Electronics & Telecomm. Department
J.D.I.E.T. Yavataml
3
Prof. Pragati D. Pawar M.E. (Electronics)
J.D.I.E.T. Yavataml
ABSTRACT
This paper presents a novel computer-aided diagnosis (CAD) technique for the early diagnosis of the Alzheimer’s disease (AD)
based on non-negative matrix factorization (NMF) and support vector machines (SVM) with bounds of confidence. For the
study and classification of functional brain images the CAD tool is designed. For this purpose, two brain image databases are
selected first one : a single photon emission computed tomography (SPECT) database and second is positron emission
tomography (PET) images, both of them containing data for both Alzheimer’s disease (AD) patients and healthy controls. The
Fisher discriminant ratio (FDR) and nonnegative matrix factorization (NMF) are used for feature selection and extraction of
the most relevant features. The NMF-transformed sets of data are classified by means of a SVM-based classifier with bounds of
confidence for decision. The proposed NMF-SVM method gives classification accuracy up to 91% with high sensitivity and
specificity rates.
1.INTRODUCTION
ALZHEIMER’s disease (AD) is the most common cause of dementia in aged people and affects more than 30 million
individuals worldwide. Dementia refers to a progressive deterioration of thinking abilities severe enough to interfere
with social, occupational and intellectual functions[2]. The particular evolution of AD patients and their increasing
dependence on the close affective environment provokes an important social repercussion. In the next 50 years the
prevalence of AD is expected to triple. Functional imaging modalities are often used with the aim of achieving early
diagnosis, although this early diagnosis remains as a demanding task. Emission computed tomography images have
been widely employed in biomedical research and clinical medicine during the last decade. These emission-based
functional images reproduce a map of physiological functions providing information about physiological phenomena
and their location in the body. In this work, two different modalities are used for brain image acquisition: Positron
Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT). Both techniques are noninvasive, which produce a three-dimensional image of functional processes in the body, such as blood perfusion or
glucose metabolism, by means of emitting radio nuclides (tracers)[3]. In both techniques, PET and SPECT, all these
detected emissions are processed and a three-dimensional image of the region under study is obtained. Computer-aided
diagnosis (CAD), are procedures in medicine that assist doctors in the interpretation of medical images. It helpto
scandigital images and is a relatively younginterdisciplinary technology combiningelements of artificialintelligence and
digital image processing with radiological image processing. CAD systems help physicians by either identifying
patterns that might have been overlooked or by providing a road map of suspicious areas. There are two different
approaches for designing CAD systems for the Alzheimer’s disease diagnosis- The first approach is the statistical
parametric mapping (SPM) tool,developed specifically to studyby comparing groups of images. The second approach is
based on the analysis of the images, the analysis of regions of interest (ROI), feature extraction and posterior
classification in different classes by means of some discriminative functions. To overcome the problems of above two
approaches a supervised learning-based CAD system applied to functional imaging consists of several important stages:
1) Functional image acquisition and normalization
2) Feature selection and reduction
3) Classification, with a train and test strategy.
Along with the adequate description of its forming techniques for feature selection , extraction and for posterior
classification,design and proper validation of a CAD tool for AD diagnosis, is the main issue in this work.
Volume 3, Issue 12, December 2014
Page 268
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
2.FUNCTIONAL IMAGE ACQUISITION AND NORMALIZATION
Each voxel of a brain functional 3-D image contains information of the corresponding brain point. However, not all the
voxels have the same level ofrelevance in terms of discrimination between groups of subjects. In this case, two groups
of subjects are defined: Alzheimer’s disease patients, label as AD, and subjects not affected by this disease, label as
NOR.
2.1Intensity Normalization
Previous to any kind of feature selection, the data sets have to be normalized in intensity in order to be able to compare
images according to their voxel normalized intensity levels. Regarding the intensity normalization, the normalization
to the maximum intensity level may introduce problems in some images that can have peak intensity values due to
noise. Thus, these images are badly normalized as the normalization is based on wrong noisy voxels. In this work, in
order to avoid these possible normalization errors, the intensity normalizationis based on the mean value of a group of
voxels with the highest intensity values. Thus, the mean value of the 0.1% voxels with the highest intensity levels is
selected for the intensity normalization.
3.FEATURE SELECTION AND REDUCTION
An initial feature selection based on discrimination capability is typically selected, obtaining a vector of discriminant
voxels for each participant. In addition, the selected discriminant voxel vectors can be projected onto a different
subspace. This subspace is chosen so that only a few variables represent the most discriminant features of each patient
imagesin each database. These steps are described below.
3.1 Fisher Discriminant Ratio for Feature Selection
The Fisher discriminant ratio (FDR) criterion is characterized by its separation. For the two-class case, it may be
defined as follows:
(μ1 ─ μ2)2
FDR= ────
(1)
σ1 2 ─ σ2 2
Whereμi and σi2 denote the class mean value and variance for each input variable, respectively.For a given variable, the
ratio value grows as the difference of the mean values of each two classes increases or the cumulative scattering in each
class decreases. In the case of thefunctional images, the voxels that satisfy a particular FDR threshold level are selected
as the most discriminative variables[4].
3.2 Nonnegative Matrix Factorization for Feature Reduction
Nonnegative matrix factorization (NMF) is a technique for finding linear representations of nonnegative data, being a
useful decomposition tool for multivariate data.This technique is especially suitable for nonnegative data sets such as
functional images and for the PET and SPECT brain images. Given a nonnegative data matrix A, NMF finds an
approximate factorization V ≈WH into nonnegative matrices W and H. Nonnegative matrix factorization is a linear,
nonnegative approximate data representation where the original database V=[V1,V2,……,VM] (N by M elements)
which consists of measurements (profiles) of nonnegative scalar variables, is approximated by a nonnegative matrix
product, as given in
V≈WH
(2)
where the matrix W=[W1, W2,……,WK] has dimension N×K, and the matrixH=[H1, H2,...,HM] has dimension K×M.
Thus, each element of matrix is decomposed, as shown in,
k
Vnm= ∑ WnkHkm ------(3)
k=1
An appropriate decision on the value of K is critical in practice, but the choice of K is very often problem dependent. In
most cases, however, K is usually chosensuch that K<<min (M, N) in which case WH can be seen as compressed form
of data in [5]. This property yields a reduced-variable matrix H that represents A in terms of the NMF basis W.After
NMF factorization, the data contained in (K by M elements) can be considered a transformed database with lower rank
(K), than the original database V.The relative error (%) of the factorization can be computed by means of the
comparison of matrix Vand the approximation WH. The minimum number of vectors (K) in the NMF basis is selected
so that a predefined level of relative error is not exceeded.
3.2.1 Factorization Rule
Given the data matrix, the optimal choice of matrices W and H are defined to be those nonnegative matrices that
minimize the reconstruction error between V and WH. A variety of error functions have been proposed some of the
most useful are given below, in (4) and (5)
1
Err1 = —― ||V―WH||2
NM
Volume 3, Issue 12, December 2014
Page 269
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
where (4) is known as the Frobenius norm (reduction of the Euclidean distance), and (5) as the Kullback–Leibler
divergence among others. The NMF process is subject to minimize the Err. The alternating least squares algorithms
(ALS) is choose due to their fast convergence and lower iteration requirements for NMF in this work. In this work only
the linear case is considered and NMF is selected, due to the simplicity of the proposed factorization and the
preservation of a linear relation between the original space of features and the new one. Fig.1 provides the first three
vectors of the NMF basisW, in the form of 2-D images, derived from one of the data sets used in this work, along with
one particular transaxial slice and the one obtained from the NMF projection.The transaxial slices provided in Fig.1 are
oriented from posterior (top) to anterior.
Figure.1 NMF projection for the sametrans-axial slice of all the SPECT database subjects.(a) NMF eigenvectors for
k=1, 2, 3. (b) One example subject slice. (c) Its NMF reconstruction
4.SVM BASED CLASSIFIER WITH BOUNDS OF CONFIDENCE
4.1SVM Background
Support vector machine (SVM) is a widely used technique for pattern recognition and classification in a variety of
applications for its ability for detecting patterns in experimental databases. SVM has become an essential machinelearning method for the detection and classification of particular patterns in medical images[6].SVM techniques consist
of two separate steps: first of all a given set of binary label training data is used for training; then new unlabeled data
can be classified according to the learned behaviour. SVM separates a given set of binary label training data by means
of a hyperplane that is maximally distant from the two possible classes (in our particular case, NOR an AD classes).
The objective is to build a function with the training data F, as expressed in (6), able to properly classify new
unclassified data
F: RN ― {±1}
(6)
The training data are formed by p different profiles each one containing N variables, together with their proper label
(NOR or AD)Thus, the training database can be expressed as
Sp= [(x1, x2,.....,xN). y]p = x .y ]p (7)
wherexN are the variables of the profile p and y the corresponding label.Linear discriminant functions define decision
hyperplanes in the N-dimensional feature space
g(x) = wTx + w0
(8)
Volume 3, Issue 12, December 2014
Page 270
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
wherew is the weight vector that is orthogonal to the decision hyperplane and w0 is the threshold. The hyperplane is
not unique and the selection process focuses on maximizing the generalization performance of the classifier. The
vectors that define the separation hyperplaneare called support vectors (SVs)[3].
Figure.2. Example of SVP_Nand SVN_P distribution in terms of distance tohyperplane
4.2 Bounds of Confidence
SVM is based in the definition of a decision hyperplane in the N-dimensional space. The SVs are a subset of the
training dataset and are chosen so that the decision hyperplane that is defined by means of them is the best one in terms
of separability between two classes. The particular distance-like decisions in SVM make this machine learning
approach especially suitable for the selection of a security region around the decision hyperplane in which decisions
may not be adopted. This security region is defined by means of the selection of bounds of confidence for the decision.
When the conventional SVM learning approach is performed and the resulting SVM classifier is trained with a set of
training observations, the separation hyperplane of these data is defined. Ideally all the training data should be
properly classified when they are re-entered in the SVM for test and classification. The hyperplane defined by all the
SVs may not be able to properly define the class of one particular support vector. That means although the SVs help to
define the best hyperplane in terms of seperability some ofthem may not be in the right subspace defined by the
hyperplane.In most of the cases, the two classes are not completely separable and there is some kind of overlapping
between them in the space of interest. It is useful to compute the classification of all the SVs and to derive the error
probability for the proper classification of the SVs of class +1 (SVPs) and SVs of class -1(SVNs). If we denote SVP_N
as a support vector ofclass +1 wrongly labelled as -1 and SVN_P on the contrary, the error probabilities for SVPs
(PerrPos) and SVNs (PerrNeg) can be defined as follows:
NSVP_N
NSVN_P
PerrPos = ————— , PerrNeg = ————------------NSVPs
NSVNs
where NSVN_P, NSVP_N, NSVPS, and NSVNS denote the number of vectors of each group. This error probability can be
reduced if SVs which are nearer to the hyperplane are not considered. The consequence of not considering these SVs in
the SVM with bounds of confidence is that all the observations located between the discarded SV which is furthest to
the hyperplane and the hyperplane itself are not classified because they are in the security zone. Fig.2 reveals clearly
this fact.
5.IMPLEMENTATION
5.1Functional Brain Image Data Sets
In order to validate the performance and outcomes of the designed NMF-SVM based CAD tool for Alzheimer’s disease
detection, two different databases are used. The first one involves SPECT brain images, whereas the second one
consists of PET brain images. These two databases contain spatially normalized functional brain images of different
subjects. This normalization step ensures that a given voxel in one patient refers to the same brain position than the
same voxel in another patient. Then, the intensities of the functional images are normalized to the maximum
intensity[7].
5.1.1SPECT Database
SPECT is a nuclear medicine tomographic imaging technique using gamma rays. However, it is able to provide true 3D information. This information is typically presented as cross-sectional slices through the patient, but can be freely
reformatted or manipulated as required.Brain perfusion images were reconstructed from projection data by
filteredbackprojection (FBP) in combination with a Butterworth noise filter[4]. The SPECT images were labelled using
two different labels: NOR for subjects without any symptom, and AD for Alzheimer’s patients. Fig.3shows a set of
trans-axialbrain slices of the SPECT database, for one AD patient and one NOR subject.
Volume 3, Issue 12, December 2014
Page 271
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
Figure.3 SPECT transaxial brain slices, oriented from posterior (top) to anterior (bottom) in each slice, and from
ventral (top left) to dorsal (bottom right) in the complete set. (a) Man mask for NOR subjects. (b) Example AD patient.
(c) Between-subjects variability between AD and NOR groups.
5.2PET Database
PET is a functional imaging technique that produces a three-dimensional image of functional processes in the body.
This detects pairs of gamma rays emitted indirectly by a positron-emitting radionuclide (tracer), which is introduced
into the body on a biologically active molecule. Thecomputer analysis isused to construct three-dimensional images of
tracer concentration within the body. If the biologically active molecule chosen for PET is fludeoxyglucose (FDG),
analogue of glucose, the concentrations of tracer imaged will indicate tissue metabolic activity by virtue of the
regionalglucose. Fig. 4 show a set of transaxial plane brain slices of the PET database, for one AD patient and one
NOR subjectglucose.
5.3 CAD Tool Evaluation
In order to evaluate the developed NMF-SVM CAD tool in all its variations, the success rate (Acc), sensitivity (Sens),
and specificity (Spec) are obtained, these SENS and SPECT defined as:
TP
TN
SENS = ──────
, SPECT = ───────
TP + FN
TN + FP
where TP is the number of positives (AD patients correctly classified); TN is the number of true negatives (NOR
patients correctly classified); FP is the number of false positives (NORclassified as AD); FN is the number of false
negatives (AD classified as NOR).
Figure.3. PET transaxial brain slices, oriented from posterior (top) to anterior(bottom) in each slice, and from ventral
(top left) to dorsal (bottom right) in the complete set. (a) Mean mask for NOR subjects. (b) Example AD patient. (c)
Between-subjects variability between AD and NOR groups.
Volume 3, Issue 12, December 2014
Page 272
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
6.EXPERIMENTAL RESULTS AND DISCUSSION
First of all a NMF-SVM based CAD tool is developed, with linear SVM as classifier. The proposed method is later
enhanced with the addition of bounds of confidence for the classification decision. Finally a new SVM-based classifier
is evaluated.In the PET database, only NOR and AD groups are considered for the validation of the CAD system, in
order to avoid errors and the uncertainty of their actual state as AD or NOR.
6.1 Basic NMF-SVM CAD Tool
Figs.4 and 5 provide the experimental results of the basic NMF-SVM system, for a variety of K values in the NMF
projection. As it is seen, levels in the range of 80% to 90% are achieved, for both databases. These results are
considered as a reference for the modifications of the CAD tool.
Figure.4. Performance of the basic NMF-SVM CAD system with the PET database, for different K values in NMF.
Figure.5. Performance of the basic NMF-SVM CAD system with the SPECT database, for different K values in NMF.
6.2 Basic NMF-SVM Tool With Bounds of Confidence
The addition of bounds of confidence to the SVM classifier permits the improvement of the results. Even though in
this case some subjects are not classified, this approach issimilar to real-life cases: sometimes some patients are difficult
to be classified with reliability even by experts. Fig.6. show the diagnosis results of the basic NMF-SVM tool with
bounds of confidence, along with the number of unlabeled patients due to the existence of a security region where
classification decisions are not allowed
Figure.6. (a) Performance of the basic NMF-SVM CAD system with bounds of confidence in the PET database, for
different K values in NMF, (b) number ofsubjects without classification in the NMF-SVM CAD system with bounds of
confidence.
7.CONCLUSION
This paper presents a NMF-SVM based technique for computed aided diagnosis of Alzheimer’s disease. The proposed
technique is based on the combination of nonnegative matrixfactorization (NMF) for feature selection and reduction
and SVM with bounds of confidence for classification. The feature reduction step provides a reduced set of variables
representingthe original data. This feature reduction is especially suitable for machine learning techniques such as
SVM. Three different approaches for the classifier are provided, two ofthem including bounds of confidence and taking
Volume 3, Issue 12, December 2014
Page 273
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 12, December 2014
ISSN 2319 - 4847
advantage of the definition of a “security region” in the SVMhyperplane, where no decision is assumed. The validation
results of the proposed NMF-SVM method yields up to 91% classification accuracy with high sensitivity and specificity
values (upper than 85%) for both data sets.
REFERANCES
[1] P. Padilla, M. López, J. M. Górriz, J. Ramírez, D. Salas-González, I. Álvarez, NMF-SVM Based CAD Tool
Applied to Functional Brain Images for the Diagnosis of Alzheimer’s Disease, ieeetransaction on medical imaging,
vol. 31, no. 2, february 2012
[2] Aiswarya.V.S, Jemimah Simon, Diagnosis of Alzheimer’s disease in Brain Images using Pulse Coupled Neural
Network, International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075,
Volume-2, Issue-6, May 2013.
[3] P. Padilla a, J.M. Górriza,, J. Ramíreza, E.W. Langb, R. Chavesa, F. Segoviaa, M. Lópeza, D. Salas-Gonzáleza, I.
Álvareza, Analysis of SPECT brain images for the diagnosis of Alzheimer’s disease based on NMF for feature
extraction, Department of Signal Theory, Networking and Communications, University of Granada, uentenueva
s/n, Granada, Spain b CIML Group, Biophysics, University of Regensburg, Germany.
[4] I. Álvarez, J. M. Górriz, J. Ramírez, D. Salas, M. López, C. G. Puntonet, and F. Segovia, “Alzheimer’s diagnosis
using eigenbrains and support vector machines,” IET Electron. Lett., vol. 45, no. 1, pp. 165–167, Feb.2009.
[5] D. D. Lee and S. Seung, “Algorithms for non-negative matrix factorization,”Adv. Neural Inf. Process. Syst., vol.
13, pp. 556–562, 2001
[6] A. Takemura, A. Shimizu, and K. Hamamoto, “Discrimination ofbreast tumors in ultrasonic images using an
ensemble classifier basedon the adaBoost algorithm with feature selection,” IEEE Trans. Med.Imag., vol. 29, no. 3,
pp. 598–609,Mar.2010.
[7] I. Illán, J. M. Górriz, J. Ramírez, D. Salas-González, M. López, F.Segovia, R. Chaves, M. Gómez-Rio, and C.
Puntonet, “18F-FDG PETimaging analysis for computer aided Alzheimer’s diagnosis,” Inf. Sci.,vol. 181, no. 4, pp.
903–916,2011.
AUTHOR
Ms. Tejal A. Fuse persuing degree in Electronics & Telecommunication Engineering from J.D.I.E.T.
Yavatmal
Ms. Nikita D. Jayasignpure persuing degree in Electronics & Telecommunication Engineering from
J.D.I.E.T. Yavatmal
Prof. Pragati D. Pawar received the B.E. degrees in Electronics & Telecommunication Engineering
from K.I.T.S, Ramtek and M.E. in Electronics from S.G.G.S. Nanded
Volume 3, Issue 12, December 2014
Page 274